risk and adaptation: evidence from global hurricane ... · risk and adaptation: evidence from...
TRANSCRIPT
Risk and Adaptation: Evidence from GlobalHurricane Damages and Fatalities
Laura A. Bakkensen∗ and Robert O. Mendelsohn†
Abstract
We examine whether countries adapt to hurricanes. A spatially re-fined global tropical cyclone data set is created to test for adaptation.We find evidence of adaptation in most of the world by examining theeffects of income, population density, and storm frequency on damageand fatalities. In contrast, there is no evidence of adaptation to dam-age in the United States leading to a damage function which is twentytimes higher than the rest of the world. (JEL D81, O1, O2, Q54, Q56,R5)
∗University of Arizona, [email protected]†Yale University, [email protected]
1
Economists are well aware that people and firms adapt to risk. People use
smoke detectors for protection from residential fire (Dardis, 1980), seat belts
for protection from traffic accidents (Atkinson and Halvorsen, 1990), and
sunscreen for protection from ultraviolet rays (Dickie and Gerking, 1996).
But how much adaptation do individuals, firms, and governments already
undertake to cope with the risks of natural disasters? The expected annual
global damage from tropical cyclones (hurricanes) is $26 billion dollars plus
19,000 lives lost (Mendelsohn et al., 2012; CRED, 2012). Is this with or
without adaptation? If there is adaptation, how much damage and fatality
has been avoided?
In the absence of official government programs, the literature on tropical
cyclones often normalizes impacts. This implies that damage is proportional
to what is in harm’s way and increases proportionally with country-level
GDP (damage is proportional to both income and population) (Hsiang and
Narita, 2012; Nordhaus, 2010; Pielke et al., 2008; Pielke and Landsea, 1998).
This assumption is a natural extension of controlled experiments where dam-
age in a wind tunnel increases in unison with capital in harm’s way. This
also implies that adaptation is not effective in reducing damages. The liter-
ature is more ambiguous about fatalities, as there has long been evidence of
adaptation. However studies normalize fatalities, assuming fatality increases
proportionately with national population (Hsiang and Narita, 2012). But do
people, firms, and governments with property at risk and lives at stake take
no effective measures to protect their assets and themselves from catastrophic
2
events such as tropical cyclones? Or do people only protect themselves from
private risks such as fire and automobile accidents?
We test the adaptation hypothesis using two approaches. First, we es-
timate the income elasticity of damage and fatalities from tropical cyclones
around the globe. If the income elasticity of damage is unitary or if the
income elasticity of fatalities is zero, this would support the hypothesis of no
adaptation. We also calculate additional socioeconomic and cyclone elastici-
ties of impacts and compare with theoretical thresholds to test for additional
types of adaptation. Second, we compare tropical cyclone damage in the
United States to tropical cyclone damage in the rest of the world. The
United States receives an average of 4 percent of global landfalls but incurs
sixty percent of global damages1. We argue that there is no adaptation in
the United States because households, firms, and local governments are com-
pensated for economic damage from tropical cyclones by a combination of
subsidized national flood insurance, state regulations on coastal property in-
surance rates, and generous post disaster relief programs. Households, firms,
and local governments have virtually no incentive to adapt to the economic
damage from tropical cyclones in the United States. Such relief programs are
at much smaller scales in the rest of the world. The consequence of adapta-
tion can be measured by contrasting the damage from storms in the United
States with the damage in the rest of the world controlling for storm inten-
1Calculated by the authors using data from the Centre for Research on the Epidemiol-ogy of Disasters (CRED, 2012).
3
sity, population, and income. Of course, the United States does not have a
program that compensates for lives lost. So only the property damage and
not the fatalities from tropical cyclones are different in the United States.
We formally test the adaptation hypothesis by gathering spatially-explicit
data on 1,400 tropical cyclone landfalls from 1960 until 2010 that have struck
inhabited areas around the world. This represents the complete set of storms
for which damage and fatality impacts are publicly available2. We match
information about the strength of these storms as well as the income and
population density of the places where the hurricanes hit. We then regress the
observed damage and fatalities from these storms on the hurricane strength
as well as the population density and income of the affected area. We also
compute the underlying frequency of low and high intensity storm landfalls
for each area and explore to what degree prior experience affects the impacts
per storm.
We are not the first to tackle the question of adaptation. Several pa-
pers find evidence of adaptation in response to development and institutions
(Kahn, 2005; Toya and Skidmore, 2007; Kellenberg and Mobarak, 2007;
Fankhauser and McDermott, 2013), as well as the underlying risk of disas-
ter (Keefer, Neumayer, and Plumper, 2011; Schumacher and Stobel, 2011;
Hsiang and Narita, 2012). Additionally, a rich literature examines the im-
2Hsiang and Jina (2014) note there are 6,700 storms in recorded history after 1950.However, many do not make landfall and fewer still have records of damage and fatalityimpacts. We collect the complete set of storms that have publicly-available recordedimpacts.
4
pact of hurricanes on long-run economic growth (Skidmore and Toya, 2002;
Strobl, 2011; Cavallo et al., 2013; Hsiang and Jina, 2014). Cavallo and Noy
(2010) and Kousky (2012) provide informative review papers. Similar to the
existing literature, we only identify the benefits of adaptation and not the
costs or net benefits. However, previous literature has shown that at current
levels of adaptation, benefits likely far exceed costs in coastal areas (Yohe,
Neumann, and Ameden, 1995; Ng and Mendelsohn, 2005). Other studies
on hurricane effects include Hallstrom and Smith (2005) and Smith et al.
(2006).
We build upon this important work in several original ways. First, we offer
a comprehensive framework to approach the question of adaptation, extend-
ing to both damages and fatalities and covering multiple channels through
which adaptation is motivated (including level of development, population
density, storm characteristics, and underlying risk of storm landfalls). Sec-
ond, we collect the most spatially-refined global dataset to date. We include
socioeconomic data at the sub-country level and use a unit of observation at
the country-landfall, not country-year. This allows us to directly model dam-
ages and fatalities per storm, instead of modeling aggregated annual values.
It also insures that any missing data are excluded from the analysis instead
of assumed to be zero. Third, we find clear evidence that adaptation matters
and quantify the reduction in damages from adaptation. Unlike previous
work that finds adaptation only reduces long run damages and fatalities by
around 3 percent (Hsiang and Narita, 2012), we find it is greatly protective
5
and reduces impacts by more than an order of magnitude.
We find ample evidence of the benefits of adaptation to fatalities across
the world. The income elasticity to fatalities is quite negative. We also
find evidence of adaptation to economic damage in every country except the
United States. The income elasticity of damage in the rest of the world lies
between 0.6 and -2.3. All of these values are statistically significantly less
than unitary. In contrast, the income elasticity of damage in the US lies
between 1 and 1.6. The hurricane damage in the United States is about
20 times higher than the rest of the world per storm. If the United States
had the same damage coefficients as the rest of the world, the expected
annual American damage from hurricanes would be $0.8 billion instead of
the observed $15.3 billion. If the rest of the world had the same damage
coefficients as the United States, non-US global damage would be $208 billion
per year instead of the observed $10.4 billion. The results suggest that a
great deal of the potential damage from tropical cyclones has been eliminated
by adaptation, except in the United States.
1 Theory
Faced with a set of risks, individuals and governments often take steps to pro-
tect themselves. We define adaptation broadly, as any action that reduces the
expected damages or fatalities from a storm. Broad-scale improvements in
forecasting and tracking as well as advanced warning systems are included as
6
adaptation. Large infrastructural development in flood protection including
levees, river channelization, and beach nourishment are included. Further,
improvements in building codes, zoning ordinances, or individual activities
to strengthen homes, including hurricane shudders and stronger foundations,
all fall under the definition. Additional adaptive channels are not consciously
undertaken to protect against storms but are still included in our definition.
For example, if societies transition toward service-based economies through
development, thereby decreasing the supply-chain and inventory sensitivi-
ties to hurricanes, we consider this adaptive. Activities that would not fall
within our definition of adaptation include insurance, as this tool financially
protects the homeowner but does not decrease damages.
Adaptation drives a wedge between observed and potential damage and
fatalities (Brooks, 2003; Fankhauser and McDermott, 2013). To empiri-
cally identify this adaptive wedge between observed and potential losses, we
first characterize the distribution of human population and capital stock in
harm’s way. Gridded global population data are available (Dobson et al.,
2000; Bhaduri et al., 2002; CIESIN et al., 2005) but spatially-explicit census
data on the global capital stock across time are not available. Several proxy
databases exist (Nordhaus, 2006; De Bono, 2013) but the detailed variation
across space is driven mainly by population rather than income per capita
assumptions.
We follow the literature by predicting the capital stock from population
and income. We calculate the ratio between capital and per capita gross
7
domestic product (GDP per capita) to be 2.65 using 2005 country-level data
from the World Bank3. This is similar to the 2.8 value from Hallegate et
al. (2013) and the 3.1 value from Kamps (2004) but well below the 5 value
assumed by Hanson et al. (2011). Thus, the empirical evidence supports the
assumption that the capital stock scales proportionately with income and
population4. We consequently assume the per capita capital stock, K, is
K = 2.65Y
where Y is income per capita.
The potential damage per storm, PDx, is the damage expected in the
absence of adaptation. It is determined by the per capita capital stock, K,
the population struck by the storm, Pop, and the intensity of the storm, I.
Because we do not know the precise size of each storm, we proxy for the
population struck by that storm using the population density of the affected
location. We use two measures of intensity, minimum pressure and maximum
wind speed. We assume damage in the absence of adaptation to have the
following functional form:
PDx = α0Y PopIα3
Similarly, we assume that potential fatalities per storm, PFx, has the
3R2 value of 0.96.4Graphs and additional supporting evidence are available in the Appendix.
8
following functional form with respect to storm intensity, I, and population
density, Pop:
PFx = β0PopIβ3
Increases in population will lead to proportional increases in potential fa-
talities. With no adaptation, income does not enter the potential fatalities
function. People of every income are equally likely to die if nobody takes
precautions. The parameters, α and β, are assumed to be positive implying
an increase in any of the above factors are expected to increase potential
impacts, including increases in income (dPDdY
> 0), increases in population
density ( dPDdPop
> 0 and dPFdPop
> 0), and increases in storm intensity (dPDdI
> 0
and dPFdI
> 0).
We next assume that individuals choose some level of adaptation, A,
with benefit B(A) and cost C(A). Assuming that the adaptation benefit and
cost functions are well behaved, the optimal adaptation, A∗, occurs when
the marginal benefit equals the marginal cost, MB(A∗) = MC(A∗). We do
not assert that adaptation is necessarily efficient in this paper. We simply
test whether individuals, firms, and governments respond to higher levels of
benefits of adaptation by doing more adaptation. That is, we assume actors
choose some nonzero level of adaptation denoted by A1 based on the marginal
damage curve MD1 in Figure 1. With adaptation level A1, the total observed
damage equals the area of triangle A1EA3 whereas the total potential damage
(with no adaptation) is triangle 0MD1A3. The fraction of damage removed
9
to the potential damage, θ(A), is θ(A) = (0MD1EA1)/(0MD1A3). Note
that the removed damage is not the welfare gain of adaptation. The welfare
gain of adaptation A1 is the area below the MB1(A) and above MC(A)
curve, as one must subtract the adaptation cost5. Observed damage, Dx, is
the product of potential damages times the fraction of damage removed by
adaptation: Dx = θ(A) · PDx.
Figure 1: Marginal Costs and Marginal Benefits of Adaptation
Several factors can shift the MB(A) curve, from MB1(A) to MB2(A)
in Figure 1, impacting the level of potential and observed damages. The
5There terms can be equivalently defined by the following integrals:∫ A3
A1MB1(A)dA for
observed damage,∫ A3
0MB1(A)dA for potential damage, and
∫ A3
0MB1(A)dA−
∫ A3
A1MB1(A)dA∫ A3
0MB1(A)dA
for the adaptation impact θ(A).
10
marginal benefit of adaptation increases with income, population, storm in-
tensity, and underlying storm frequency (Π). Under an efficient solution, this
would also increase the equilibrium level of adaptation. However, we do not
require optimality, we simply test whether A2 > A1. That is, we test whether
adaptation increases as income, population density, or storm frequency in-
creases ( dAdY
> 0), ( dAdPop
> 0)6, ( dAdΠl
> 0) and ( dAdΠh
> 0). We specifically
examine the effect of predicted frequencies of both low (Πl) and high (Πh)
intensity storms. Incorporating potential demand shifters, we approximate
θ(A) with the following constant elasticity functional form:
θ(A) ≈ (1− γ0)Y −γ1Pop−γ2I−γ3x Π−γ4l Π−γ5
h
The γi terms equal zero if there is no adaptation. The observed damage will
have the following expression:
Dx = α0(1− γ0)Y 1−γ1Pop1−γ2Iα3−γ3x Π−γ4
l Π−γ5h
Similarly, observed fatalities, Fx, from storm x are the multiplicative product
of potential damages and adaptation, Fx = θ(A) · PFx:
Fx = β0(1− γ0)Y −γ1Pop1−γ2Iβ3−γ3x Π−γ4l Π−γ5
h
6This may be especially true if public adaptation is focused on areas with more people,but if adaptation costs increase in population, then there may be no increase in adaptation.
11
We test this proposition using both cross sectional evidence across loca-
tions as well as intertemporal variation in our panel. The study explores
whether adaptation increases as factors that would increase the potential
benefits of adaptation increase. If no adaptation is present in economic
damage and fatalities, then γi = 0 for i = {0, 1, 2, 3, 4, 5}. Whether γi > 0
is a testable hypothesis for the existence of adaptation. That is, adapta-
tion is present in economic damage to the extent that the income elasticity
and population elasticity are less than unitary (1). Adaptation would also be
evident if the historic frequency of storms lowers the damage per storm. Sim-
ilarly, adaptation is present in fatalities if the elasticity of income is negative,
the elasticity of population is less than one, or the elasticity with respect to
frequency is negative. Note however, that the potential coefficient on the
constant term and on the intensity of storms is not known and so cannot be
used to test for adaptation. Thus, relative comparisons within sample can
be made but there is not theoretical threshold based on our model.
Insurance is an important mechanism to cope with risks (Arrow, 1973;
Kunreuther, 1996). In some cases, insurance can be a substitute for adapta-
tion, especially when it is set at below actuarially-fair rates or coupled with
subsidized post-disaster aid (Kelly and Kleffner, 2003). Because the United
States has subsidized flood insurance (Michel-Kerjan, 2010), generous post
disaster compensation, and regulations on traditional insurance premiums
along the coast, individuals, firms, and local governments face very low costs
for being in harm’s way. Their insurance premiums for the additional risk
12
are near zero and they are often compensated for damage that is not in-
sured. Owners of both governmental and private capital along the coast in
the United States have no incentive to adapt to the risk of tropical cyclones.
There of course may be additional reasons why the United States is differ-
ent than the rest of the world. The US may have a stronger desire to live
along the coast and it may be wealthier. However, the study controls for
population density as well as income per capita at the county level.
By comparing the storm damage function of the United States versus the
rest of the world, one can test the assumption of the model parameters above.
We postulate that the damage function for the United States reflects potential
damage (zero abatement). The rest of the world exemplifies adaptation
since it does not offer such generous compensation programs. In contrast,
both the United States and the rest of the world should have similar fatality
functions. Another test of adaptation is therefore the difference between the
parameters of the United States damage function and the parameters of the
damage function for the rest of the world. The damage function of the
United States should resemble the potential damage function whereas the
damage function of the rest of the world resembles the adaptation damage
function. Note that we are not assuming perfectly efficient adaptation in the
rest of the world, just more extensive adaptation. Because the United States
does not compensate people for dying in a tropical cyclone, the United States
is not expected to have different fatality coefficients, just different damage
coefficients. The United States damage function also offers a good test of the
13
appropriate elasticities of the potential damage function. The United States
comparison also provides a test of the constant term and the coefficient on
storm intensity.
For all of these tests of adaptation, we are assuming that there are many
possible actors that can adapt including households, firms, and farms as well
as local and state governments. We are assuming that private actors focus
on reducing just their own damages, while governments focus on reducing
the damages to all the people in their jurisdiction. This analysis does not
distinguish who is doing the adaptation. We therefore are examining the
adaptation of both private individuals and firms as well as local govern-
ments. We do not know to what extent adaptation to economic damage is
a complement or substitute to adaptation to fatalities. Lastly, we do not
address the costs of adaptation and therefore do not calculate the net ben-
efits of adaptation. However, it is worth noting that the coastal protection
literature concerned with sea level rise finds that it is generally cheaper to
protect developed land than let it be inundated (Yohe et al., 1995; Ng and
Mendelsohn, 2005; Hunt et al., 2011; Neumann et al., 2011).
2 Empirical Strategy
Guided by the theoretical framework above, we use panel data to test for
the presence of adaptation to tropical cyclone damages and fatalities. We
14
first estimate damage and fatality functions using a log-log functional form
with cross-sectional and panel techniques7. Resulting estimated coefficients
can be interpreted as elasticities. We then test to see if these elasticities are
below theoretical thresholds (evidence of adaptation). We also test if adap-
tation levels vary across income levels by estimating respective damage and
fatality elasticities on partitioned samples including only low or high income
countries. This also reduces the potential for any strategic measurement er-
ror in the damage reports. Next, using our spatially refined data, we test for
the importance of local versus national adaptation. Finally, we compare the
elasticities of the United States to those of the rest of the world8.
We use both a cross-sectional model and error components model with
country and time fixed effects to calculate damage and fatality functions.
Cross-sectional analysis relies primarily on the variation across space to
identify parameters of interest, whereas the identifying variation for our
error components model relies on deviations from country and year aver-
ages. Broadly, cross-sectional results can shed light on long-run patterns
of adaptation (Mendelsohn et al., 1994). To the extent that some adapta-
tion changes very slowly over time, within-country and within-year variation
will not capture adaptive changes on these broader scales. However, cross-
sectional analysis may be confounded by time- and location-specific omitted
7Count data estimation and Seemingly Unrelated Regression model results are shown inthe Appendix. The results support the findings of our cross-sectional and error componentsmodels.
8See the Appendix for a detailed explanation of specification tests and explanatoryvariable choice.
15
variable bias that our error components model will subsume. Lastly, panel
data and cross-sectional results often have a different economic interpreta-
tion, as short term shocks are different than long term adaptive potential
(Timmins and Schlenker, 2009; Samuelson, 1947). Due to the strengths and
weaknesses of each technique, we present both models herein.
After specification and model selection testing presented in the Appendix,
we chose the following log-log functional form for its goodness of fit to model
damages for cyclone landfall j at time t in country i:
lnDijt = α0+α1lnYit+α2lnPopit+α3lnIijt+α4lnLijt+α5lnΠhi+α6lnΠli+αi+γt+uijt
and for fatalities:
lnFijt = β0+β1lnYit+β2lnPopit+β3lnIijt+β4lnLijt+β5lnΠhi+β6lnΠli+αi+γt+uijt
where Dij is direct economic damages and Fij is the number of fatalities.
These impacts are explained by Yit, the income per capita in country i at
the time of cyclone j; Popit, the population density; Iijt, the intensity of
cyclone j when making landfall in country i; Πli, the long-term frequency
of low intensity storms in country i; and Πhi, the long-term frequency of
high intensity storms in country i. Lij, a variable for landfall, is 1 if the
cyclone j made a direct landfall on the country i and otherwise equal to the
distance in kilometers of the storms’ closest approach. Since the variable
Lij is not present in our theoretical model, as a robustness check we also
16
drop this variable in the Appendix and find no change in the overall results.
In the error components model, we also include fixed effects for time (γt)
and country (αi). uijt is a mean-zero error term. Explanatory variables are
identical between the cross sectional and fixed effects specification except for
the year, γt, and country, αi, fixed effects which subsume the high and low
intensity cyclone frequency variables.
We estimate both functions using the Ordinary Least Squares (OLS) es-
timator. We also cluster standard errors at the country level in all specifica-
tions unless noted otherwise, to account for any within-country correlation
across error term observations9. While we include near misses in our main
result. In the Appendix we present results with near misses dropped, thereby
allowing our empirical model to exactly replicate our theoretical model. The
results do not change with the inclusion of near misses.
Unlike previous literature that aggregates up to the country-year level
(Hsiang and Narita, 2012; Neumayer, 2012; Noy, 2009; Kahn, 2005), one
major difference in our analysis is that our unit of observation is a country-
landfall (a storm striking a country). This means that a country suffering
from three storms in a year will be treated as three observations. There are
several advantages. First, this ensures that any missing storms or missing
data on storm impacts are not treated as zero, which could bias estimated
coefficients. Second, we directly model damages and fatalities at the storm
9Ferreira et al. (2013) note the importance of country-clustered standard errors forcross-country disaster analyses.
17
level. Thus, we can use more spatially refined data, including individual
storm characteristics at their point of landfall as well as sub-national socioe-
conomic controls. Lastly, we do not normalize cyclone impacts by population
or GDP, which would imply no adaptation. Using country level GDP or pop-
ulation to normalize for what is in harm’s way is problematic as each storm
hits a different place within a country with different characteristics.
In the Appendix, we present count data technique results for fatalities,
estimating semi-log regressions with the Negative Binomial estimator. We
test for and find evidence of over-dispersion in the data, implying that the
Negative Binomial estimator is preferred to the Poisson. We find that OLS
is appropriate when modeling the (log of) damages, as this variables follows
a normal distribution rather than a Poisson or Negative Binomial distribu-
tion. Fixed effects negative binomial results are included, but should be
interpreted with caution as there is still some debate in the literature as
to proper implementation of the fixed effect controls and what is actually
being subsumed by different implementations (Greene, 2007). The results
support the findings of our cross-sectional and fixed effects results. We also
use the Seemingly Unrelated Regression Model to potentially leverage effi-
ciency gains over OLS through exploitation of any correlation in the error
terms. However, we do not find this changes the main results and present
the results in the Appendix.
In addition to these main results, we do some robustness checks to test
selected sub-samples across: levels of development, spatial scales (national
18
versus local), and between the US and the rest of the world.
2.1 Data
For the empirical analysis, we build an original dataset of more than 1,400
storm landfalls around the Earth from 1960 to 2010 totaling almost $0.75
trillion in damages10 and approximately 400,000 lives lost. We begin our
analysis in 1960, coinciding with the start of satellites used for storm ob-
servations. Before this period, storms were incidentally observed by ships
and coastal communities, or found by aircraft flying long-range patterns in
search of disturbances (HRD, 2014). Thus, we have less confidence in the
accuracy of storm data and human impacts before this period. Hsiang and
Narita (2014) correctly note that 6,700 storms have been recorded by hu-
mans since 1950, but many of these storms do not make landfall and fewer
still can be linked with direct economic damages or human fatalities. Thus,
our dataset represents the full record of storms during our time period that
can be matched with publicly available damages and fatalities. We present
summary statistics in the Appendix.
Historical cyclone landfall damage and fatalities are from the EM-DAT
Emergency Disaster Database (CRED, 2012) and Nordhaus (2010) and are
matched with tropical cyclone characteristics compiled by NOAA IBTrACS
v03r03, U.S. Navy Tropical Cyclone Reports, and Nordhaus (2010). Both
maximum wind speed and minimum sea level pressure are tested as proxies
10All dollar values in this paper are in terms of real 2010 $USD.
19
for cyclone intensity. Additionally, we include the Power Dissipation Index
and Accumulated Cyclone Energy Index as cyclone intensity proxies in the
Appendix.
Ideally, analyses of damages and fatalities would control for the exact
population and capital impacted by the storm. However, the spatial extent
of a storm is not recorded by IBTrACS for most storms11. Thus, many studies
use country-level socioeconomic variables as proxies. We collect both coun-
try and sub-country data. We collect country-level population density and
per capita income data come from the Penn World Table v7.01, USDA ERS
International Macroeconomic Data, the CIA World Factbook, and Columbia
CIESIN’s Gridded Population of the World v3. In addition to national data
collected annually for the globe, we also collect sub-national, county-level
population density and income per capita data for six large countries (Aus-
tralia, China, India, Japan, Philippines, United States, and Mexico at the
state-level) using official census records. Storms to these six countries rep-
resent approximately 60 percent of our sample. The remaining countries
are small- to medium-sized whose national statistical more closely represent
the local levels. This allows us to more accurately assess the socioeconomic
conditions at landfall. Note, too, that by using income per capita (instead
of national GDP) and population density (instead of total population) we
can change spatial scale without impacting the overall level of damages. For
11The radius of maximum winds is recorded for a limited number of storms in theNorthern Atlantic.
20
example, smaller countries should have the same damage function as larger
countries. We test the importance of using country versus sub-country data
in the Results section. We also test both market exchange rate and purchas-
ing power parity definitions for income per capita and present our results in
the Appendix.
Since the size of the storm, its spatial extent, varies by storms but is
not accurately recorded, one cannot control for size in this study. We con-
sequently do not know the aggregate population or capital in harm’s way
of each storm. We approximate what is in harm’s way by using population
density and per capita income in the nearest counties to landfall. This allows
a much closer fit between where storms land and what is impacted than using
national GDP or population. Also, since our variables are per capita or per
square mile, we are also able to carefully test the benefits of using national
versus sub-national data in estimating damages.
Finally, a hurricane generator is used to predict the long-term frequency
for low and high intensity storm landfalls for each location. We turn to sim-
ulation data because the historical record of storm tracks is heterogeneous
in quality across time and space, especially before the development of the
Dvorjac technique that greatly improved accuracy in estimating hurricane
strength and the large-scale satellite deployment in the 1970s (Velden et
al., 2006). A total of 68,000 simulated cyclone tracks generated by Kerry
Emanuel are used to predict the frequencies by location (Emanuel et al.
2006; Mendelsohn et al., 2012). For the purposes of this analysis, low inten-
21
sity storms have 10-minute sustained maximum wind speeds that rank them
between a tropical depression and Category 3 strength (34 to 115 knots).
High intensity storms include all Category 4 and 5 storms (greater than 115
knots), based on wind speed (NHC, 2012). We present the summary statis-
tics of the sample in our Appendix. All together, 87 countries are struck by
tropical cyclones and are represented. Only observed landfalls are included
in the database, locations with no storms are omitted from our analysis.
With any data, measurement error is possible. In this analysis, there may
be measurement error with our estimates of damage (EM-DAT), income, and
cyclone intensity. All are addressed herein. The damage and fatality data
are likely the largest source of potential classical measurement error and
even strategic reporting bias. The bias introduced by strategic reporting
could impact accuracy in both directions: countries may try to under-report
damage to appear more capable, while other countries may try to over-report
damage to encourage international aid, relief, and sympathy. This could be
particularly true for lower income countries. Classical measurement error
will cause no bias in the regression coefficients but underreporting could bias
impacts downward. EM-DAT, the data provider, takes care to collect data
from multiple sources and verify the accuracy of the reports. If countries con-
sistently misreport data, then it would be observed during cross-verification
by EM-DAT from reports by the UN, World Bank, Red Cross, and other
organizations. EM-DAT prioritizes data from the most trusted sources. In
addition, we control for potential strategic reporting through selective sub-
22
sample regressions based on income, assuming that within income groups,
countries will not systematically differ in their incentives to miss-report. We
present our findings in the Results section and also the Appendix. We do
not find evidence that strategic reporting changes our results.
Income and GDP records may also have measurement error in report-
ing and estimation. We test a variety of data sources, including the Penn
World Table and USDA ERS International Macroeconomic Data, and both
market exchange rate and purchasing power parity definitions of GDP. We
also use our low versus high income partitioned regression results to address
potential measurement error concerns. Assuming the measurement error is
not consistent across data sources or within levels of development, similar
empirical results give us confidence that potential measurement error is not
a large factor biasing estimates.
Finally, measurement error is possible in the storm intensity record. Sci-
entific ability to accurately describe storm intensity has greatly improved in
the 1970s and 1980s with large scale satellite deployment and technique im-
provements (Velden et al., 2006). However, it is likely that minimum sea level
pressure is measured with greater accuracy than maximum wind speed (Gray
et al., 1991; and see our discussion in the Appendix and the Results section).
We minimize any potential measurement error by using multiple proxies for
storm intensity including maximum wind speed, barometric pressure, PDI,
and ACE. We find that minimum sea level pressure explains both damage
and fatalities more accurately than the alternative intensity measures.
23
Another important issue is that of selection of observations into our anal-
ysis sample. EM-DAT is arguably one of the best sources for global natural
disaster data available and verification of data quality is an important step
in their entry procedures (Tschoegl et al., 2006; Guha et al., 2002). However,
not all historical cyclone landfalls are included in the EM-DAT database and
not all cyclones in the database have a record of both damage and fatalities.
EM-DAT censors low impact storms with minimum damage and fatality cri-
terion12. However, the definition of a tropical cyclone itself censors storms
below a critical intensity. It is possible this censorship of small, low damage
storms affects the estimated parameters but this censorship is not expected
to have a large effect on the overall results because few fatalities and little
damage are caused by low intensity storms.
3 Results
This section presents our main results using cross-sectional and fixed effects
specifications. We find our results robust to alternative specifications, func-
tional forms, and additional sensitivity analyses. Our robustness checks are
presented in the Results section and detailed in the Appendix. Addition-
ally, in the Appendix, we drop near misses, presenting only the results for
which the distance from landfall equals zero. This empirical specification is
12A cyclone must meet at least one of the following criterion to be included in EM-DAT: 1) 10 or more fatalities, 2) 100 or more people affected, 3) a declaration of a stateof emergency, or 4) a call for international assistance (CRED, 2012).
24
identical to our theoretical model.
3.1 Fatalities
Table 1 shows the regression results for our fatality function using all coun-
tries. Columns 1, 2, and 3 are cross-sectional regressions. Column 1 presents
a basic regression. Column 2 decomposes the underlying cyclone frequency
into low, ΠL, and high, ΠH , intensity storms. Column 3 uses maximum wind
speed instead of minimum sea level pressure as a proxy for storm intensity.
Columns 4 and 5 add a year fixed effect. Columns 6 and 7 add a country
fixed effect. Note that the t-statistic on observed coefficients may be used
to test if estimated elasticities are statistically different from zero. We use
the F-test to test if relevant elasticities are statistically different from one.
The signs of the estimated elasticities are as we expected, with fatalities ris-
ing with lower minimum sea level pressure and higher maximum sustained
wind speed, I. Fatalities decrease as the distance from the eye of the storm
increases.
25
Tab
le1:
Evid
ence
ofA
dap
tati
onto
Fat
alit
ies
Dep
end
ent
Var
iab
le:
Log
Fat
alit
ies
(1)
(2)
(3)
(4)
(5)
(6)
(7)
Reg
ress
ion
Bas
eΠ
Sp
lit
Win
dY
ear
FE
Yea
rF
E,
Win
dC
ountr
yF
EC
ou
ntr
yF
E,
Win
d
Ln
Inco
me
Per
Cap
ita
(Y)
-0.6
18**
*-0
.651
***
-0.6
53**
*-0
.618
***
-0.6
11**
*-0
.218
***
-0.1
35*
(0.0
834)
(0.0
886)
(0.0
871)
(0.0
868)
(0.0
863)
(0.0
738)
(0.0
684)
Ln
Pop
ula
tion
Den
sity
(Pop
)0.
146*
0.13
2*0.
106
0.14
5*0.
121
0.22
8***
0.2
24***
(0.0
786)
(0.0
772)
(0.0
817)
(0.0
870)
(0.0
910)
(0.0
509)
(0.0
694)
Ln
Inte
nsi
ty(Ix
Pre
ssu
re)
-9.1
89**
*-9
.429
***
-8.2
70**
*-1
0.54
***
(2.7
77)
(2.7
91)
(2.9
05)
(2.3
55)
Ln
Inte
nsi
ty(Ix
Win
dS
pee
d)
0.57
1***
0.38
4**
0.6
48***
(0.1
45)
(0.1
75)
(0.1
39)
Ln
Fre
qu
ency
All
(Π)
0.07
83*
(0.0
416)
Ln
Fre
qu
ency
Low
(ΠL
)0.
257*
*0.
248*
*0.
279*
**0.
273*
**(0
.103
)(0
.104
)(0
.099
6)(0
.099
6)L
nF
requ
ency
Hig
h(Π
H)
-0.1
18*
-0.1
20*
-0.1
35**
-0.1
31**
(0.0
673)
(0.0
670)
(0.0
653)
(0.0
643)
Ln
Lan
dfa
llD
ista
nce
(L)
-0.1
62**
*-0
.158
***
-0.1
57**
*-0
.149
***
-0.1
51**
*-0
.141
***
-0.1
39***
(0.0
231)
(0.0
227)
(0.0
222)
(0.0
227)
(0.0
232)
(0.0
216)
(0.0
219)
Con
stan
t69
.97*
**70
.86*
**3.
966*
**63
.35*
**4.
411*
**77
.76*
**1.8
41*
(19.
53)
(19.
67)
(1.1
40)
(20.
49)
(0.9
46)
(16.
36)
(0.9
86)
Yea
rF
EN
NN
YY
YY
Cou
ntr
yF
EN
NN
NN
YY
Ob
serv
atio
ns
1,00
61,
006
995
1,00
699
51,
020
1,0
08
R-s
qu
ared
0.23
50.
243
0.22
90.
297
0.29
00.
234
0.2
41
Not
e:**
*p<
0.01
,**
p<
0.05
,*
p<
0.1
All
spec
ifica
tion
sh
ave
stan
dar
der
rors
clu
ster
edat
the
cou
ntr
yle
vel.
26
Using our theoretical thresholds, we find strong evidence of adaptation
to fatalities. The income elasticity with respect to fatalities is less than
zero, β1 < 0, for all specifications, lying between -0.618 and -0.135. This
income elasticity of fatalities is consistent with the income elasticity of the
value of statistical life, found at the global meta-level to be between 0.5 to
0.6 (Viscusi and Aldy, 2003; Viscusi, 1993). We reject the null hypothesis
that the income elasticity is equal to zero for all specifications, and reject at
the 93% confidence level for the more conservative specification in column 7
where the elasticity is closest to zero. We also find evidence of adaptation
to fatalities with respect to population density, β2 < 1. Using the F-test, we
find that the estimated elasticities are all less than one at the 99% confidence
level. The population density coefficient implies that places with more people
suffer more fatalities. However, the coefficient also implies that the fatalities
per person are lower in more dense places. That is, from a personal point of
view, cities are safer than more rural areas. This may be a conscious urban
policy of adaptation for example due to urban evacuation plans or this result
could simply be a consequence of constructing dense and sturdy structures
in cities (Lindell et al., 2011; Whitehead, 2003).
We find a divided result for the underlying storm frequency. The coeffi-
cient on the frequency of high intensity storms elasticity is negative, β5 < 0,
implying places that are often hit by powerful storms have taken precautions.
Keefer et al. (2001) find similar results with lower fatalities to earthquakes
in areas hit more frequently. However, we find the opposite result for the fre-
27
quency of low intensity storms, (ΠL), as the estimated elasticities are greater
than zero, β4 > 0. This finding is significant at the 95% confidence level
in Column 2 through 5. Although this analysis does not specify the mal-
adaptive mechanism, one possible explanation is that individuals suffer from
warning fatigue. Frequent weak storms pose small risks that do not war-
rant dramatic responses. With frequent false alarms, people may stop taking
even modest precautions. Lastly, since people react differently to low and
high intensity storms, a variable characterizing overall frequency of storms,
Π, hides this dichotomous relationship. Thus, we caution the use of a single
variable to characterize underlying hazard risk commonly used in the litera-
ture (Fankhauser and McDermott, 2013; Hsiang and Narita, 2012; Neumayer
et al., 2013; Schumacher and Strobl, 2011; Keefer et al., 2011).
28
Tab
le2:
Evid
ence
ofA
dap
tati
onto
Dam
ages
Dep
end
ent
Var
iab
le:
Log
Dam
ages
(1)
(2)
(3)
(4)
(5)
(6)
(7)
Reg
ress
ion
Bas
eΠ
Spli
tW
ind
Yea
rF
EY
ear
FE
,W
ind
Cou
ntr
yF
EC
ou
ntr
yF
E,
Win
d
Ln
Inco
me
Per
Cap
ita
(Y)
0.44
7**
0.42
0**
0.36
4**
0.40
3**
0.35
3**
0.02
70.1
23
(0.1
96)
(0.1
85)
(0.1
75)
-0.1
87(0
.175
)(0
.157
)(0
.169)
Ln
Pop
ula
tion
Den
sity
(Pop
)0.
074
0.05
7-0
.001
0.06
1-0
.034
-0.0
52-0
.303**
(0.1
28)
(0.1
26)
(0.1
54)
(0.1
26)
(0.1
54)
(0.2
07)
(0.1
33)
Ln
Inte
nsi
ty(Ix
Pre
ssu
re)
-29.
49**
*-2
9.94
***
-28.
40**
*-3
4.35
***
(6.2
69)
(6.0
61)
(5.2
88)
(7.3
08)
Ln
Inte
nsi
ty(Ix
Win
dS
pee
d)
1.86
9***
1.73
8***
1.9
97***
(0.3
83)
(0.4
12)
(0.4
89)
Ln
Fre
qu
ency
All
(Π)
-0.0
454
(0.1
01)
Ln
Fre
qu
ency
Low
(ΠL
)0.
169
0.23
9*0.
224
0.27
9**
(0.1
40)
(0.1
41)
(0.1
39)
(0.1
39)
Ln
Fre
qu
ency
Hig
h(Π
H)
-0.1
44-0
.170
*-0
.171
*-0
.189
*-0
.094
4-0
.097
8-0
.090
-0.0
957
Ln
Lan
dfa
llD
ista
nce
(L)
-0.4
14**
*-0
.413
***
-0.3
64**
*-0
.393
***
-0.3
49**
*-0
.360
***
-0.3
17***
(0.0
528)
(0.0
517)
(0.0
606)
(0.0
523)
(0.0
560)
(0.0
577)
(0.0
627)
Con
stan
t21
7.2*
**21
9.3*
**5.
879*
*20
8.9*
**6.
559*
*25
4.7*
**10.9
5***
(42.
63)
(41.
31)
(2.7
01)
(35.
44)
(2.9
28)
(49.
76)
(3.1
35)
Yea
rF
EN
NN
YY
YY
Cou
ntr
yF
EN
NN
NN
YY
Ob
serv
atio
ns
844
844
832
844
832
856
843
R-s
qu
ared
0.22
30.
227
0.21
20.
282
0.27
00.
246
0.2
33
Not
e:**
*p<
0.01
,**
p<
0.05
,*
p<
0.1
Sta
nd
ard
erro
rsar
ecl
ust
ered
atth
eco
untr
yle
vel.
29
3.2 Damage
Table 2 shows the results of the damage regressions using data from all coun-
tries. The column specifications are identical to those of Table 1. Damage
increases with the intensity of the storm13 and decreases with distance from
the eye. We find clear evidence of adaptation in the income elasticity with
respect to damage. The income elasticity varies from .03 to .45. The es-
timated income elasticities are all significantly less than one, α1 < 1. We
perform an F-test and reject (at the 99 percent confidence level) the unitary
income elasticity of damages assumed by the previous literature (Hsiang and
Narita, 2012; Nordhaus, 2010; Pielke et al., 2008; and Pielke and Landsea,
1998).
The coefficient on population density varies between -0.3 and 0.07. These
values are all significantly less than one, α2 < 1. As population density
increases, damages do not increase. This result indicates damage per person
falls in urban areas. Again this result may be due to conscious policies to
adapt urban areas to storms or it may simply be an incidental result of more
sturdy structures in urban areas.
Lastly, we find the elasticity of damage with respect to storm intensity
to be lower than past literature. For example, the elasticity of minimum
pressure is -29 to -34 whereas previous studies using data from the United
States found values of -86 (Mendelsohn et al., 2012). The elasticity of damage
13Recall that minimum sea level pressure has an inverse relationship with intensity; astronger storm has a lower pressure reading.
30
with respect to maximum sustained winds is from 1.7 to 2 which is much
closer to the traditional literature which found damage increases with the
second or third power of wind speed (Emanuel, 2005; Bell et al., 2000; Pielke
and Landsea, 1999). In contrast, the empirical results from US data imply
much higher elasticities of between 5 and 9 (Nordhaus, 2010; Mendelsohn et
al., 2012).
Based on the Vuong (1989), AIC, and BIC tests, we prefer the use of min-
imum sea level pressure over wind speed14. We also use the PDI and ACE as
additional proxies for storm intensity and present the results in the Appendix.
One explanation of the superiority of minimum pressure as a measure of in-
tensity is that wind speed may be measured with greater error than minimum
sea level pressure (Gray et al., 1991). Additionally, wind speed calculation
techniques have changed over time without good documentation whereas
minimum pressure reading techniques have remained consistent over time
(Emanuel, 2013). Maximum wind speed is calculated differently throughout
the world, reflecting 1-, 3-, or 10 minute sustained maximum wind speeds.
As there is no deterministic relationship between these different measures,
statistical averages must be used to convert them, leading measurements to
diverge from the true values. Finally, some wind speed estimates across the
world have been derived statistically from pressure readings whereas other
measures have relied on rules of thumb making it difficult to track the source
14We also test using both pressure and wind speed, but both variables become insignif-icant due to high multicollinearity.
31
of wind data (NRL, 1998). Thus, we recommend the use of minimum sea
level pressure readings to be utilized for tropical cyclone damage and fatality
research.
32
Tab
le3:
Adap
tati
onto
Fat
alit
ies:
Low
and
Hig
hIn
com
eC
ountr
ies
Dep
enden
tV
aria
ble
:L
ogF
atal
itie
s(1
)(2
)(3
)(4
)(5
)(6
)In
com
eL
evel
<$6
,500
<$6
,500
<$6
,500
>$2
0,00
0>
$20,
000
>$2
0,00
0R
egre
ssio
nB
ase
Dec
ade
FE
Cou
ntr
yF
EB
ase
Dec
ade
FE
Cou
ntr
yF
E
Ln
Inco
me
Per
Cap
ita
(Y)
-0.4
47**
*-0
.406
***
-0.0
01-2
.277
***
-2.7
48**
*-1
.751
**(0
.146
)(0
.131
)(0
.119
)(0
.369
)(0
.511
)(0
.621
)L
nP
opula
tion
Den
sity
(Pop
)0.
361*
**0.
372*
**0.
291*
**0.
134
0.14
60.
150*
*(0
.074
5)(0
.078
6)(0
.086
8)(0
.105
)(0
.089
4)(0
.068
7)L
nIn
tensi
ty(Ix
Pre
ssure
)-1
3.59
***
-11.
68**
*-1
3.01
***
-7.2
66-9
.116
*-2
.857
(2.6
92)
(2.6
16)
(2.8
94)
(5.4
69)
(5.0
91)
(3.3
30)
Ln
Lan
dfa
llD
ista
nce
(L)
-0.2
00**
*-0
.184
***
-0.1
60**
*-0
.146
***
-0.1
56**
*-0
.163
***
(0.0
204)
(0.0
213)
(0.0
223)
(0.0
394)
(0.0
324)
(0.0
472)
Con
stan
t98
.84*
**85
.69*
**92
.97*
**74
.47*
92.0
5**
39.5
1(1
8.93
)(1
8.16
)(2
0.46
)(3
8.65
)(3
6.80
)(2
4.39
)
Dec
ade
FE
NY
YN
YY
Cou
ntr
yF
EN
NY
NN
YO
bse
rvat
ions
579
579
579
152
152
152
R-s
quar
ed0.
184
0.20
90.
175
0.17
00.
210
0.13
1N
ote:
***
p<
0.01
,**
p<
0.05
,*
p<
0.1
All
spec
ifica
tion
shav
est
andar
der
rors
clust
ered
atth
eco
untr
yle
vel.
33
Tab
le4:
Adap
tati
onto
Dam
ages
:L
owan
dH
igh
Inco
me
Cou
ntr
ies
Dep
enden
tV
aria
ble
:L
ogD
amag
es(1
)(2
)(3
)(4
)(5
)(6
)In
com
eL
evel
<$6
,500
<$6
,500
<$6
,500
>$2
0,00
0>
$20,
000
>$2
0,00
0R
egre
ssio
nB
ase
Dec
ade
FE
Cou
ntr
yF
EB
ase
Dec
ade
FE
Cou
ntr
yF
E
Ln
Inco
me
Per
Cap
ita
(Y)
0.51
3***
0.60
6***
0.34
7*-1
.721
*-2
.311
*-2
.158
**(0
.178
)(0
.194
)(0
.204
)(0
.994
)(1
.118
)(0
.793
)L
nP
opula
tion
Den
sity
(Pop
)0.
405*
*0.
410*
*0.
003
0.00
8-0
.023
70.
510*
**(0
.166
)(0
.156
)(0
.076
3)(0
.175
)(0
.180
)(0
.147
)L
nIn
tensi
ty(Ix
Pre
ssure
)-2
4.33
***
-22.
77**
*-2
4.13
***
-37.
74**
*-3
8.20
***
-43.
23**
*(4
.132
)(3
.874
)(3
.421
)(1
2.15
)(1
3.31
)(1
1.44
)L
nL
andfa
llD
ista
nce
(L)
-0.3
94**
*-0
.398
***
-0.3
82**
*-0
.540
***
-0.5
43**
*-0
.659
***
(0.0
501)
(0.0
601)
(0.0
508)
(0.1
61)
(0.1
64)
(0.1
16)
Con
stan
t17
9.0*
**16
8.5*
**18
1.9*
**29
5.4*
**30
4.6*
**33
4.9*
**(2
8.33
)(2
6.66
)(2
3.73
)(8
0.96
)(9
1.61
)(7
5.92
)
Dec
ade
FE
NY
YN
YY
Cou
ntr
yF
EN
NY
NN
YO
bse
rvat
ions
414
414
414
145
145
145
R-s
quar
ed0.
230
0.25
10.
201
0.25
60.
278
0.28
4N
ote:
***
p<
0.01
,**
p<
0.05
,*
p<
0.1
All
spec
ifica
tion
shav
est
andar
der
rors
clust
ered
atth
eco
untr
yle
vel.
34
3.3 Adaptation Across Income Levels
One hypothesis that has been raised with respect to adaptation is that adap-
tive capacity rises with income. We test this hypothesis in Tables 3 and 4
by examining whether the income elasticity of damage and fatalities is lower
for higher income locations. We create sub-samples of the data for low in-
come (<$6,500) and high income (>$20,000) locations. We then estimate
separate regressions on each subsample. The United States is dropped as an
outlier in this analysis. Table 3 reveals the results for fatalities. The columns
vary depending upon the income of the locations and the use of fixed effects.
Columns 1 and 4 are OLS regressions, columns 2 and 5 have decade fixed ef-
fects and columns 3 and 6 have both time and country fixed effects. Standard
errors are clustered at the country level. To check the validity of the clustered
standard errors for subsample regressions with fewer than fifty bins, we also
calculate the coefficient p-values using wild bootstrapping as described by
Cameron et al. (2008) and implemented in Stata with Caskey (2013). The
significance of the results do not change. We use locations and not countries
for this analysis. The included high (low) income locations come mainly from
highly (lesser) developed countries and also wealthy urban centers (lower in-
come rural areas) in developing and lesser developed countries, controlling
also for population density. Therefore, differences are not driven by national
policies but location-specific differences between wealthy versus poorer areas.
Low income locations have an income elasticity with respect to fatalities
from 0 to -0.4 whereas high income locations have an income elasticity from
35
-1.8 to -2.7. These results provide strong support for the theory that people
adapt to prevent fatalities. Adaptation increases with income. The high
income location elasticities are statistically different from the elasticities of
low income locations at the 99% confidence level. These very negative income
elasticities of fatalities imply a much higher relationship between income and
value of life compared to the literature (Viscusi and Aldy, 2003; Viscusi,
1993). The remaining coefficients of the fatality model are not different for
the two subsamples.
Table 4 presents the damage results for low and high income locations.
The columns in each damage regression are identical to those in Table 3 for
fatalities. The income elasticity of damage for low income locations varies
between 0.35 and 0.61 whereas the income elasticity varies between -1.7 and
-2.3 for high income locations. All included countries show signs of adapta-
tion to economic damage, and once again the results imply that adaptation
increases rapidly with income, even overcoming the scale effect of more in
harm’s way. The damage income elasticity results are similar to the pro-
jections from an environmental Kuznets curve, with damages first increasing
and then decreasing with income (Shafik, 1994). The estimated population,
intensity, and distance coefficients are not statistically different between low
and high income countries.
36
3.4 Adaptation Across Space
Using the collected national and sub-national data, we then test adaptation
across levels of spatial scale. This allows us to estimate the gains from
using more nuanced spatial data. Sub-national data gives us insight into
local adaptation and the differences between adaptation in urban versus rural
areas. National socioeconomic data allows us to identify the broad differences
across levels of economic development and federal policies. Across space and
time, the county-level income per capita records are highly correlated with
the national levels (ρ = 0.94), while the population density is less correlated
across spatial scales (ρ = 0.75). Thus, within countries, the urban versus
rural population variation dominates the economic inequality across space.
In Table 5, we formally test the difference between national and sub-
national adaptation. Using the 60 percent of our data that has both national
and sub-national socioeconomic data, we estimate our impact functions us-
ing exclusively national-level data and then the same observations using sub-
national measures. The other 40 percent of the data includes many small
countries with less variation between local and national socioeconomic vari-
ables. Columns 1-4 show our fatality functions, while Columns 5-8 report
our estimated damage functions. Odd columns present county-level results
and even columns represent the equivalent regressions using national-level
socioeconomic data. Lastly, we drop the United States in Columns 3, 4, 7,
and 8. We highlight again that our population and income variables are nor-
malized by area and population, thus allowing us to navigate up and down
37
spatial scales without impacting the aggregate level of damages calculated.
For example, one landfall may be defined by a national per capita average in-
come of $15,000 and a local per capita average income of $20,000, but would
result in similar levels of estimated damages. These tests also allay fears of
our use of sub-national data in our main empirical results.
We find overall that both county- and national-level data have high ex-
planatory power for both damages and fatalities. Although there are few
statistically significant differences, the interpretation of results can yield po-
tentially important findings. Data from the national level shows higher rates
of adaptation, with slightly larger income elasticities of fatality (-0.87 versus
-0.91) and damage (-0.35 vs -0.50). This highlights the importance of the
overall level of development of a country allowing for better national-level
adaptation, potentially including efforts in advanced storm warning systems
or better coordination of post-disaster aid. Similarly, for the population den-
sity elasticity, damages show the strongest differences. Denser nations have
lower (more negative) elasticities than their more rural counterparts. Simi-
larly, within countries, rural areas have higher aggregate damages than urban
areas. This striking finding highlights the importance of public adaptation in
urban areas and also the potential difficulties protecting sparsely populated
areas.
38
Tab
le5:
Adap
tati
onA
cros
sSpat
ial
Sca
les
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Dep
enden
tV
aria
ble
(Ln)
Fat
alit
ies
Fat
alit
ies
Fat
alit
ies
Fat
alit
ies
Dam
ages
Dam
ages
Dam
ages
Dam
ages
Ln
Inco
me
Per
Cap
ita
(Cou
nty
)-0
.737
***
-0.8
74**
*0.
0589
-0.3
53**
*(0
.066
5)(0
.076
9)(0
.104
)(0
.122
)L
nIn
com
eP
erC
apit
a(N
atio
nal
)-0
.784
***
-0.9
05**
*-0
.298
***
-0.5
04**
*(0
.063
6)(0
.073
9)(0
.104
)(0
.111
)L
nP
opula
tion
Den
sity
(Cou
nty
)0.
0672
0.07
27-0
.265
***
-0.2
60**
*(0
.049
9)(0
.054
7)(0
.074
3)(0
.076
5)L
nP
opula
tion
Den
sity
(Nat
ional
)-0
.087
1*-0
.001
54-0
.727
***
-0.6
20**
*(0
.052
6)(0
.058
9)(0
.085
8)(0
.083
4)L
nIn
tensi
ty(P
ress
ure
)-1
6.03
***
-15.
26**
*-1
5.75
***
-15.
09**
*-2
8.26
***
-33.
46**
*-2
1.70
***
-24.
40**
*(2
.593
)(2
.557
)(2
.747
)(2
.703
)(4
.041
)(3
.799
)(4
.190
)(3
.934
)L
nL
andfa
llD
ista
nce
-0.1
71**
*-0
.168
***
-0.1
57**
*-0
.159
***
-0.4
99**
*-0
.408
***
-0.4
98**
*-0
.434
***
(0.0
323)
(0.0
323)
(0.0
337)
(0.0
336)
(0.0
673)
(0.0
648)
(0.0
634)
(0.0
608)
Ln
Cou
nt
Low
0.24
60.
556*
*0.
409
0.51
6**
2.68
0***
3.53
7***
3.05
4***
3.40
7***
(0.3
11)
(0.2
59)
(0.3
24)
(0.2
61)
(0.5
34)
(0.4
46)
(0.5
24)
(0.4
08)
Ln
Cou
nt
Hig
h-0
.252
**-0
.332
***
-0.3
69**
*-0
.413
***
-1.4
57**
*-1
.756
***
-1.7
27**
*-1
.824
***
(0.1
19)
(0.1
07)
(0.1
24)
(0.1
10)
(0.2
09)
(0.1
84)
(0.2
06)
(0.1
70)
Con
stan
t11
8.8*
**11
1.6*
**11
7.1*
**11
2.3*
**19
4.0*
**22
7.2*
**15
0.1*
**16
8.3*
**(1
8.37
)(1
7.94
)(1
9.49
)(1
8.99
)(2
8.59
)(2
6.32
)(2
9.46
)(2
7.36
)
Obse
rvat
ions
574
574
518
518
497
497
389
389
Spat
ial
Sca
leC
ounty
Nat
ional
Cou
nty
Nat
ional
Cou
nty
Nat
ional
Cou
nty
Nat
ional
USA
Incl
uded
?Y
YN
NY
YN
NR
-squar
ed0.
259
0.26
90.
267
0.27
80.
306
0.37
80.
342
0.41
3Sta
ndar
der
rors
inpar
enth
eses
**p<
0.01
,**
p<
0.05
,*
p<
0.1
39
3.5 The United States
The United States has been dropped as an outlier in earlier global studies
of tropical cyclone impacts (Hsiang and Narita, 2012). Here, we utilize the
economic damage results from the United States as an example of potential
damage (no adaptation). We first test to see if the United States is different
from the rest of the world, and then decompose the rest of the world into
highly and lesser developed nations to test for additional differences.
We first test to see if the estimated damage and fatality coefficients for the
United States are different than the rest of the world using both an F-test and
Chow Breakpoint test. We first conduct an F-test by including an indicator
variable for the United States (1 if USA, 0 otherwise). We then interact
this variable with the included variables, to allow for slope variation. We
then test if the sum of all US estimated coefficients is statistically different
from non-US coefficients. For damages, we calculate an F-statistic of 30.90,
rejecting the null hypothesis that the US and rest of the world coefficients
are equal to each other at more than the 0.1 percent level. For fatalities, we
calculate an F-statistic of 2.26 and fail to reject the null hypothesis that the
US and the rest of the world are the same at the 10 percent level.
We also conduct the Chow Breakpoint test to determine if the estimated
coefficients of subsamples of the entire sample are statistically different from
each other. It does not, however, test if the sample is optimally partitioned.
We perform the Chow breakpoint test on the USA and non-USA subsample
regressions. For the damage regressions, we calculate a chi-squared statistic
40
of 60.32, rejecting the null hypothesis at the 0.1 percent level that the co-
efficients of the explanatory variables are the same across sub-groups of the
data. For the fatalities regressions, we calculate a chi-squared statistic of
3.54 and fail to reject the null hypothesis at even the 40 percent level. Thus,
we conclude that the United States has a similar fatality function as the rest
of the world but a very different damage function.
Table 6 compares the regression results estimated by subsamples from the
United States, OECD countries excluding the United States, and non-OECD
countries. Results for both minimum pressure and wind speed are presented.
We find that the income elasticity of the United States varies between 1.1
and 1.6 which is consistent with the zero adaptation case. In contrast, the
income elasticity for the remaining countries in the OECD lies between -
0.5 and -0.6 and for the non-OECD countries between 0.2 and 0.3 which is
consistent with adaptation. The population density coefficient is negative for
the United States which is not consistent with the zero adaptation case. We
consequently conclude that the coefficient on population density is not a good
test of adaptation. A storm that strikes American cities causes less damage
than a storm that strikes rural areas. The discrepancy is more marked than
in any other country. The effect of storm intensity is higher in the United
States than the rest of the world. Damages escalate rapidly with intensity
in the United States. The elasticity with respect to minimum pressure is
-85 for the United States but -34 for the rest of the OECD and -24 for
non-OECD countries. Similar differences exist for wind speed. Finally, the
41
constant term is much higher for the United States implying that all storms
cause more damage.
How much higher would damages be across the globe if the rest of the
world did not adapt, and how much lower would damages be in the US if it
adapted like the rest of the world? We use our estimated impact function
to calculate these counter-factual scenarios. We use the US coefficients as
the example of zero adaptation and the rest of the world as the example of
adaptation. If the United States had the same damage coefficients as the
rest of the world, the annual tropical cyclone damages in the United States
would average $0.8 billion instead of the current $15.3 billion. If the rest
of the world had the same damage coefficients as the US, non-US global
damages would be $208 billion per year instead of the current $10.4 billion.
There is a 20 fold difference between the estimated damage function for the
US and the rest of the world. Thus, we find that adaptation matters greatly
in driving a large wedge between potential and observed impacts.
42
Tab
le6:
Unit
edSta
tes
Dam
ages
Dep
enden
tV
aria
ble
:L
ogD
amag
es(1
)(2
)(3
)(4
)(5
)(6
)C
ountr
ies
USA
USA
OE
CD
OE
CD
non
-OE
CD
non
-OE
CD
&non
-USA
&non
-USA
Reg
ress
ion
Bas
eW
ind
Bas
eW
ind
Bas
eW
ind
Ln
Inco
me
per
Cap
ita
1.14
8**
1.63
6***
-0.6
24-0
.459
0.28
5***
0.22
9**
(0.5
48)
(0.5
55)
(0.3
95)
(0.4
24)
(0.0
986)
(0.0
995)
Ln
Pop
ula
tion
Den
sity
-0.3
00-0
.342
0.29
8***
0.30
9**
0.09
800.
0677
(0.2
66)
(0.2
84)
(0.0
707)
(0.1
31)
(0.0
869)
(0.0
858)
Ln
MSL
P-8
4.75
***
-34.
35**
-23.
70**
*(7
.969
)(1
4.03
)(3
.312
)L
nM
axim
um
Win
d5.
069*
**2.
005
1.42
5***
(0.6
22)
(1.4
50)
(0.2
39)
Ln
Lan
dfa
llD
ista
nce
-0.1
35-0
.033
9-0
.690
***
-0.6
80**
*-0
.351
***
-0.3
22**
*(0
.300
)(0
.196
)(0
.144
)(0
.149
)(0
.042
7)(0
.043
4)C
onst
ant
592.
1***
-17.
07**
260.
0***
13.8
8*17
7.9*
**9.
737*
**(5
4.80
)(6
.796
)(9
7.12
)(7
.678
)(2
2.85
)(1
.261
)
Obse
rvat
ions
108
110
9581
653
652
R-s
quar
ed0.
498
0.44
60.
334
0.31
50.
171
0.15
5N
ote:
***
p<
0.01
,**
p<
0.05
,*
p<
0.1.
43
3.6 Robustness
We present our robustness results in the Appendix. We first provide evidence
of our capital formation assumption, showing that capital scales linearly with
changes in income and population. We also present our sensitivity analysis.
Our main empirical results are robust to alternative variables, functional
forms, and additional sensitivity analyses. We carefully test the shape of the
damage and fatality functions, including linear, log-linear, quadratic, and
cubic specifications. We test the impact of data definitions and additional
proxy variables through the use of both market exchange rate and purchas-
ing power parity income and GDP per capita. We also test different proxies
for storm intensity including maximum wind speed and minimum sea level
pressure, as well as both the Power Dissipation and Accumulated Cyclone
Energy indices. An additional robustness check drops “near misses”, or hur-
ricanes that do not directly make landfall in a country, presenting only the
results for which the distance from landfall equals zero. This empirical spec-
ification is identical to our theoretical model. We find that across all of these
specifications, our main results hold.
We also test different estimators, including the negative binomial esti-
mator and the Seemingly Unrelated Regressions (SUR) model. Since we
use identical regressors in both our equations, there is no efficiency gain in
the SUR model relative to the OLS model (Greene, 2003). We test the ap-
plicability of a cross-equation coefficient restriction, but find the estimated
coefficients in the damages and fatalities equations to be statistically differ-
44
ent from each other, thereby negating the motivation of imposing equality.
This makes theoretical sense; storm and human factors may impact damage
and fatalities in different and independent ways. We also provide additional
income bins and estimate income elasticities of damage and fatalities across
twelve levels of development. We lastly present results comparing fatalities
in the United States to that of the rest of the world. Across all of our ro-
bustness checks, we find additional confirmation of the conclusions from our
main empirical results.
4 Conclusion
This paper develops a theory of adaptation to tropical cyclone damages and
fatalities in order to test for the presence of adaptation. We add to the
literature by constructing a new, larger, and more spatially-explicit historical
dataset of more than 1,400 storms, matching cyclone landfall impacts with
spatially-refined socioeconomic and cyclonic characteristics. A set of multiple
regressions are then estimated with this new dataset to test for adaptation.
Two types of tests are undertaken. First, we look at economic damage and
explore whether the elasticity of income and to a lesser extent the elasticity of
population density is unitary. We also look at fatalities and explore whether
the elasticity of income is negative and whether the elasticity of population
is less than unitary. We also test if there is a negative relationship between
impacts and the underlying storm frequency. We decompose these main
45
findings into adaptation across level of development and spatial scale. We
find clear evidence of adaptation in most of these tests. The coefficient of
population density may not be a good test of adaptation. Nonetheless, the
damages and especially the fatalities are much less than one would expect if
no adaptation measures were being undertaken.
We also use economic damage in the United States as a counterfactual.
State but mostly federal policies compensate potential victims of hurricanes.
As a result, there is little private or local government incentive to adapt in the
United States. Compensation mechanisms in the rest of the world are much
smaller. Comparing the multiple regressions using just United States data,
other OECD countries, and non-OECD countries reveals the United States
has a unique damage function. The income elasticity of damage is unitary
or higher, the elasticity with respect to storm intensity is much higher, and
the constant is higher. If the United States had the damage function of the
rest of the world, expected damages from hurricanes would be twenty times
smaller. If the rest of the world had a damage function like the United States,
damages would be twenty times higher. Adaptation to tropical cyclones is
clearly very important.
Why is the United States an outlier? One major difference between the
United States and the rest of the world is the role of public policy in shap-
ing the incentives for coastal inhabitants to undertake risk. The incentive
to adapt to tropical cyclones has been virtually eliminated in the United
States. Coastal property is almost completely compensated for any risk from
46
hurricanes by several policies. Many states limit how high insurance rates
can climb for risky coastal areas effectively subsidizing high income house-
holds along the coast (Kousky, 2011). The National Flood Insurance Pro-
gram charges insurance premiums that are well below what they must pay,
especially when tropical cyclones strike. The U.S. Government Accountabil-
ity Office finds that historical program payouts exceed premiums by $30.4
billion (GAO, 2013). Post-disaster aid is funded through general tax rev-
enues across the country instead of being paid by premiums (Krutilla, 1966;
Kousky, 2010). The expectation of post disaster aid reduces adaptation
(Kelly and Kleffner, 2003). All of these policies encourage individuals to live
in more risky areas and take few precautions to protect property. Some of the
most rapidly developing areas in the United States are coastal, with limited
incentive to retreat to safer locations and no incentive to invest in physical
protection.
This research reveals adaptation to tropical cyclones is ongoing in most
of the world and it dramatically reduces damage and fatalities. However, the
study does not provide critical details about this adaptation. How much of
the adaptation is being done by private actors and how much by local, state,
and federal governments? How much adaptation is in hard structures such as
barriers and how much is just rational land use planning? What are the in-
centives to live by the sea relative to inland in the United States versus other
countries? Very little is known about the distribution of damages within a
tropical cyclone. How much of the damage is concentrated in low elevation
47
sites along the shore? How much is due to storm surge, high winds, or fresh
water flooding? More spatially detailed measures of storm characteristics,
what is in harm’s way, and damage are needed.
The research suggests several policy insights. We find that well-intentioned
compensation of victims through public programs decreases the incentive to
adapt. Specifically, federal programs have eliminated the incentive for pri-
vate individuals and firms as well as local and state governments to adapt to
reduce damage per storm. This leads to over-capitalization in high risk areas
that drive up aggregate observed damage and to the absence of adaptation
measures. The public insurance programs in the United States need reform.
Premiums need to cover the outlays for insurance to work properly. Fair
insurance premiums provide a useful signal to households, firms, and farms
informing them where risks are high and low. When public policies prevent
premiums from reflecting the true underlying risk, they eliminate this signal
causing private individuals and firms to make poor choices. Although a na-
tional post disaster compensation program serves a valuable compassionate
role, such programs can be paid for by assessing premiums on local and state
governments consistent with long term payouts in each jurisdiction. Fair pre-
miums provide a valuable incentive for governments to manage rather than
ignore these risks.
Second, the results suggest that there may be room to improve the public
communication of cyclone risks. Places with high intensity storms appear to
have taken measures to reduce fatalities per storm. Yet fatalities per storm
48
increase in places with many low intensity storms. It is unclear why fatali-
ties are higher in places with more frequent low intensity storms. Researchers
should explore whether public warnings can be improved to eliminate this
effect. Fatalities have fallen over time in most countries again suggesting an
effective adaptation program. However, the problem is not yet completely
solved. For example, 77 percent of global fatalities occurred in just two
countries, Myanmar and Bangladesh, over the last several decades15. It ap-
pears there are still at least two countries that can significantly improve their
overall performance.
Lastly, the results strongly suggest that economic development helps in-
crease adaptation to natural disasters. The income elasticity of damage is
less than one in every country except the United States. As countries de-
velop, the damage from tropical cyclones will be a smaller component of
their income. High income countries actually have lower aggregate damage.
There is also strong evidence that per capita damage is much lower in urban
areas. To the extent that development is both increasing incomes and urban-
ization, these factors will help reduce the future burden of tropical cyclones
on the economy. It is also quite clear that development is leading to fewer
fatalities. Both urbanization and rising incomes is cutting fatalities rapidly.
Development is a key component of adaptation to natural disasters.
15Calculated by the authors using data from CRED (2012).
49
References
Arrow, Kenneth J. 1974. Optimal insurance and generalized deductibles.
Scandinavian Actuarial Journal 1974 1–42.
Atkinson, Scott E and Robert Halvorsen. 1990. The valuation of risks to life:
evidence from the market for automobiles. The Review of Economics and
Statistics 72 133–136.
Bell, Gerald D, Michael S Halpert, Russell C Schnell, R Wayne Higgins, Jay
Lawrimore, Vernon E Kousky, Richard Tinker, Wasila Thiaw, Muthuvel
Chelliah, and Anthony Artusa. 2000. Climate assessment for 1999. Bulletin
of the American Meteorological Society 81 s1–s50.
Bhaduri, Budhendra, Edward Bright, Phillip Coleman, and Jerome Dobson.
2002. Landscan. Geoinformatics 5 34–37.
Brooks, Nick. 2003. Vulnerability, risk and adaptation: A conceptual frame-
work. Tyndall Centre for Climate Change Research Working Paper 38 1–
16.
Cameron, A Colin, Jonah B Gelbach, and Douglas L Miller. 2008. Bootstrap-
based improvements for inference with clustered errors. The Review of
Economics and Statistics 90 414–427.
Caskey, Judson. 2013. Stata file cgmwildboot.ado. Available online at https :
//webspace.utexas.edu/jc2279/www/data.html.
50
Cavallo, Eduardo, Sebastian Galiani, Ilan Noy, and Juan Pantano. 2013.
Catastrophic natural disasters and economic growth. Review of Economics
and Statistics 95 1549–1561.
Center for International Earth Science Information Network, CIESIN and
Centro Internacional de Agricultura Tropical. 2005. Gridded population of
the world, version 3. CIESIN, Columbia University.
CRED, Centre for Research on the Epidemiology of Disasters. 2012. Em-
dat: The ofda/cred international disaster database, universite catholique
de louvain, brussels (belgium).
Dardis, Rachel. 1980. The value of a life: new evidence from the marketplace.
The American Economic Review 70 1077–1082.
De Bono, Andrea. 2013. Global exposure database for gar 2013.
UNEP/GRID-Geneva Working Paper.
Dickie, Mark and Shelby Gerking. 1996. Formation of risk beliefs, joint pro-
duction and willingness to pay to avoid skin cancer. The Review of Eco-
nomics and Statistics 78 451–463.
Dobson, Jerome E, Edward A Bright, Phillip R Coleman, Richard C Durfee,
and Brian A Worley. 2000. Landscan: a global population database for
estimating populations at risk. Photogrammetric engineering and remote
sensing 66 849–857.
51
Emanuel, Kerry. 2005. Increasing destructiveness of tropical cyclones over
the past 30 years. Nature 436 686–688.
Emanuel, Kerry. 2013. Anthropogenic effects on tropical cyclone activity.
Available online at http : //wind.mit.edu/ emanuel/anthro2.htm.
Emanuel, Kerry, Sai Ravela, Emmanuel Vivant, and Camille Risi. 2006. A
statistical deterministic approach to hurricane risk assessment. Bulletin of
the American Meteorological Society 87.
Fankhauser, Samuel and Thomas KJ McDermott. 2013. Understanding the
adaptation deficit: why are poor countries more vulnerable to climate
events than rich countries? Grantham Research Institute on Climate
Change and the Environment Working Paper No. 134.
Government Accountability Office, GAO. 2013. High risk report:
National flood insurance program. Available online at http :
//www.gao.gov/highrisk/nationalf loodinsurance.
Gray, William, Charles Neumann, and Tedl Tsuisui. 1991. Assessment of the
role of aircraft reconnaissance on tropical cyclone analysis and forecasting.
American Meteorological Society, Bulletin 72 1867–1883.
Greene, William. 2007. Fixed and random effects models for count data. NYU
Working Paper No. EC-07-16.
Greene, William H. 2003. Econometric analysis. Pearson Education.
52
Guha-Sapir, Debby and Regina Below. 2002. The Quality and Accuracy of
Disaster Data: A Comparative Analyses of Three Global Data Sets. World
Bank, Disaster Management Facility, ProVention Consortium.
Gutierrez, Carlos M, Cynthia A Glassman, J Steven Landefeld, and Rose-
mary D Marcuss. 2007. A primer on gdp and the national income and
product accounts. US Department of Commerce, September.
Hallegatte, Stephane, Colin Green, Robert J Nicholls, and Jan Corfee-
Morlot. 2013. Future flood losses in major coastal cities. Nature climate
change 3 802–806.
Hallstrom, Daniel G and V Kerry Smith. 2005. Market responses to hurri-
canes. Journal of Environmental Economics and Management 50 541–561.
Hanson, Susan, Robert Nicholls, Nicola Ranger, Stephane Hallegatte, Jan
Corfee-Morlot, Celine Herweijer, and Jean Chateau. 2011. A global ranking
of port cities with high exposure to climate extremes. Climatic change 104
89–111.
Hsiang, Solomon M and Amir S Jina. 2014. The causal effect of environmental
catastrophe on long-run economic growth: Evidence from 6,700 cyclones.
Technical report National Bureau of Economic Research.
Hsiang, Solomon M and Daiju Narita. 2012. Adaptation to cyclone risk:
Evidence from the global cross-section. Climate Change Economics 3.
53
Hunt, Alistair and Paul Watkiss. 2011. Climate change impacts and adapta-
tion in cities: a review of the literature. Climatic Change 104 13–49.
Hurricane Research Division, HRD. 2014. History:
The stormfury era. Available online at http :
//www.aoml.noaa.gov/hrd/abouthrd/stormfuryera.html.
Kahn, Matthew. 2005. The death toll from natural disasters: The role of in-
come, geography, and institutions. The Review of Economics and Statistics
87 271–284.
Kamps, Christophe. 2004. New Estimates of Government Net Capital Stocks
for 22 OECD Countries 1960-2001 (EPub). 4-67 International Monetary
Fund.
Keefer, Philip, Eric Neumayer, and Thomas Plumper. 2011. Earthquake
propensity and the politics of mortality prevention. World Development
39 1530–1541.
Kellenberg, Derek and Ahmed Mushfiq Mobarak. 2007. Does rising income
increase or decrease damages risk from natural disasters? Journal of Urban
Economics 63 788–802.
Kelly, Mary and Anne Kleffner. 2003. Optimal loss mitigation and contract
design. Journal of Risk and Insurance 70 53–72.
Kousky, Carolyn. 2010. Understanding the demand for flood insurance. Nat-
ural Hazards Review 12 96–110.
54
Kousky, Carolyn. 2011. Managing natural catastrophe risk: State insurance
programs in the united states. Review of Environmental Economics and
Policy 5 153–171.
Kousky, Carolyn. 2012. Informing climate adaptation: A review of the eco-
nomic costs of natural disasters, their determinants, and risk reduction
options. Resources for the Future Discussion Paper.
Krutilla, John V. 1966. An economic approach to coping with flood damage.
Water Resources Research 2 183–190.
Kunreuther, Howard. 1996. Mitigating disaster losses through insurance.
Journal of risk and Uncertainty 12 171–187.
Lindell, Michael K, Jung Eun Kang, and Carla S Prater. 2011. The logistics
of household hurricane evacuation. Natural hazards 58 1093–1109.
Mendelsohn, Robert, Kerry Emanuel, Shun Chonabayashi, and Laura
Bakkensen. 2012. The impact of climate change on global tropical cyclone
damage. Nature Climate Change 2 205–209.
Mendelsohn, Robert, William D Nordhaus, and Daigee Shaw. 1994. The im-
pact of global warming on agriculture: a ricardian analysis. The American
Economic Review 753–771.
Michel-Kerjan, Erwann O. 2010. Catastrophe economics: the national flood
insurance program. The Journal of Economic Perspectives 165–186.
55
National Hurricane Center, NHC. 2012. Saffir-simpson hurricane wind scale.
Available online at www.nhc.noaa.gov/aboutsshws.php.
National Research Laboratory, NRL. 1998. Tropical cyclone
forecasters’ reference guide. Available online at http :
//www.nrlmry.navy.mil/ chu/tropcycl.htm.
Neumann, James, Daniel Hudgens, John Herter, and Jeremy Martinich. 2011.
The economics of adaptation along developed coastlines. Wiley Interdisci-
plinary Reviews: Climate Change 2 89–98.
Ng, Wei-Shiuen and Robert Mendelsohn. 2005. The impact of sea level rise
on singapore. Environment and Development Economics 10 201–215.
Nordhaus, William. 2010. The economics of hurricanes and implications of
global warming. Climate Change Economics 1 1–20.
Nordhaus, William D. 2006. Geography and macroeconomics: New data
and new findings. Proceedings of the National Academy of Sciences of the
United States of America 103 3510–3517.
Noy, Ilan. 2009. The macroeconomic consequences of disasters. Journal of
Development Economics 88 221–231.
Pielke Jr, Roger, Joel Gratz, Christopher Landsea, Douglas Collins, M.A.
Saunders, and Rade Musulin. 2008. Normalized hurricane damage in the
united states: 1900–2005. Natural Hazards Review 9 29–42.
56
Pielke Jr, Roger A and Christopher N Landsea. 1999. La nina, el nino and
atlantic hurricane damages in the united states. Bulletin of the American
Meteorological Society 80 2027–2033.
Pielke Jr, Roger A and Christopher W Landsea. 1998. Normalized hurricane
damages in the united states: 1925-95. Weather and Forecasting 13 621–
631.
Samuelson, Paul A. 1947. Foundations of economic analysis. Harvard Uni-
versity Press, Cambridge.
Schumacher, Ingmar and Eric Strobl. 2011. Economic development and losses
due to natural disasters: The role of hazard exposure. Ecological Economics
72 97–105.
Shafik, Nemat. 1994. Economic development and environmental quality: an
econometric analysis. Oxford Economic Papers 757–773.
Skidmore, Mark and Hideki Toya. 2002. Do natural disasters promote long-
run growth? Economic Inquiry 40 664–687.
Smith, V Kerry, Jared C Carbone, Jaren C Pope, Daniel G Hallstrom, and
Michael E Darden. 2006. Adjusting to natural disasters. Journal of Risk
and Uncertainty 33 37–54.
Strobl, Eric. 2011. The economic growth impact of hurricanes: evidence from
us coastal counties. Review of Economics and Statistics 93 575–589.
57
Timmins, Christopher and Wolfram Schlenker. 2009. Reduced-form versus
structural modeling in environmental and resource economics. Annu. Rev.
Resour. Econ. 1 351–380.
Toya, Hideki and Mark Skidmore. 2007. Economic development and the im-
pacts of natural disasters. Economic Letters 94 20–25.
Tschoegl, Liz, Regina Below, and Debby Guha-Sapir. 2006. An analyt-
ical review of selected data sets on natural disasters and impacts. In
UNDP/CRED Workshop on Improving Compilation of Reliable Data on
Disaster Occurrence and Impact. World Wide Web Address: http://www.
emdat. be/Documents/Publications/TschoeglDataSetsReview. pdf.
Velden, Christopher, Bruce Harper, Frank Wells, John L Beven, Ray Zehr,
Timothy Olander, Max Mayfield, Charles Chip Guard, Mark Lander,
Roger Edson, et al.. 2006. The dvorak tropical cyclone intensity estimation
technique: A satellite-based method that has endured for over 30 years.
Bulletin of the American Meteorological Society 87 1195–1210.
Viscusi, Kip. 1993. The value of risks to life and health. Journal of economic
literature 1912–1946.
Viscusi, Kip and Joseph Aldy. 2003. The value of a statistical life: a criti-
cal review of market estimates throughout the world. Journal of risk and
uncertainty 27 5–76.
58
Vuong, Quang. 1989. Likelihood ratio tests for model selection and non-
nested hypotheses. Econometrica: Journal of the Econometric Society
307–333.
Whitehead, John C. 2003. One million dollars per mile? the opportunity costs
of hurricane evacuation. Ocean & coastal management 46 1069–1083.
Yohe, Gary, James Neumann, and Holly Ameden. 1995. Assessing the eco-
nomic cost of greenhouse-induced sea level rise: methods and application
in support of a national survey. Journal of Environmental Economics and
Management 29 S78–S97.
59